AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

CoLT: Teaching Multi-Modal Models to Think with Chain of Latent Thoughts

Lianyu Hu, shengqian qin, Zeqin Liao, Qing Guo, Liang Wan, Wei Feng, Yang Liu

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 2 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace direct answering with three latent thought vectorsQwen3-VL-8B · 8-benchmark averageaccuracy: 69.5 -> 79.1Improved
replace textual CoT with latent thoughtsQwen3-VL-8B · 8-benchmark averageaccuracy: 75.7 -> 79.1Improved
replace SIM-CoT latent reasoning with CoLTQwen3-VL-8B · 8-benchmark averageaccuracy: 74.0 -> 79.1Improved
add backward decoding loss to forward + internal losses (ablation)CoLT · SeedBench, MMStar, AI2D, MMTaccuracy: 70.2 -> 74.8Improved
add forward decoding loss to backward + internal losses (ablation)CoLT · SeedBench, MMStar, AI2D, MMTaccuracy: 71.3 -> 74.8Improved
add internal supervision loss to forward + backward losses (ablation)CoLT · SeedBench, MMStar, AI2D, MMTaccuracy: 71.9 -> 74.8Improved
add forward decoding loss alone (ablation)CoLT · SeedBench, MMStar, AI2D, MMTaccuracy: 63.6 -> 68.0Improved
increase latent steps K from 1 to 3 (ablation)CoLT · SeedBench, MMStar, AI2D, MMTaccuracy: 68.9 -> 74.8Improved
increase latent steps K from 3 to 8 (ablation)CoLT · SeedBench, MMStar, AI2D, MMTaccuracy: 74.8 -> 73.4Hurt
scale external decoder from Qwen3-0.6B to Qwen3-8B (ablation)CoLT · SeedBench, MMStar, AI2D, MMTaccuracy: 74.8 -> 75.4No clear effect
train with K=3 and test at K=6 (ablation)CoLT · SeedBenchaccuracy: 77.5 -> 76.8No clear effect
replace text CoT with CoLT under 30% occlusionQwen3-VL-8B · MMStar occlusion noiseaccuracy degradation: -13.8 -> -9.0Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy